Analysis
ChatGPT for Healthcare adds computer-use agents: what the HIPAA BAA covers, and what to set first
1AYM is an OpenAI Select Partner.
Published . Vendor facts checked on against each vendor’s own pages, listed at the end. Plans, prices and availability change. Neither OpenAI nor Anthropic has reviewed this analysis.
The short version
OpenAI published two posts on 16 November 2026, one on new features in ChatGPT for Healthcare and one on computer-use agents for healthcare teams. We read both on 17 November 2026, with OpenAI's Help Center pages on HIPAA eligible products and on ChatGPT for Healthcare. None of the four pages gives a price.
- What OpenAI announced
- On 16 November 2026 OpenAI added computer-use agents, improved trusted clinical search and more than 20 healthcare plugins. ChatGPT for Healthcare customers get them today, subject to workspace settings.
- BAA and controls
- OpenAI's eligibility page lists which settings its BAA covers. Browser use and network access for Work in the cloud are outside it, and the page does not name the new agents.
What ChatGPT for Healthcare added on 16 November 2026, and who gets it
What was announced
OpenAI announced three things for ChatGPT for Healthcare on 16 November 2026. They are advances in computer-use agents, improved trusted clinical search, and more than 20 new healthcare plugins. ChatGPT for Healthcare is OpenAI's enterprise version of ChatGPT for clinicians, administrators and researchers, run in what OpenAI calls a governed workspace.
A computer-use agent is an AI agent that works a computer as a person would. It reads the screen, clicks, types and moves between applications. OpenAI says the agents carry out multi-step work across existing healthcare software. They move information between systems that do not connect, such as the electronic health record (EHR), hospital policies, scheduling systems and payer portals.
OpenAI's worked example is an MRI. When a clinician recommends one, the details an insurer needs may already be in the record. A team member still copies them into a separate portal for the prior authorisation request. Prior authorisation is the insurer's approval before it covers some treatments and scans.
In OpenAI's demonstration, run on synthetic patient data, the agent reviews visit notes, physical therapy records and earlier imaging in the EHR. It drafts an MRI order, an encounter note and a reply to the patient for the clinician to review. It then prepares the prior authorisation request in the insurer's portal for an authorisation specialist to review.
OpenAI names CVS Health, Cigna Healthcare, Abridge and Ambience among the organisations working with the agents. CVS Health uses them to process referrals, send documents between systems and support specialist consults. OpenAI says one CVS agent cleared a months-long document backlog over a weekend. Another handles the system steps for tens of thousands of consults each day.
Cigna Healthcare's case managers used to move between more than 10 separate clinical and administrative systems to gather context. OpenAI says computer use now brings that information together in seconds. It quotes Cigna Healthcare's CIO as saying the work went from concept to deployment in under eight weeks. Abridge is still developing its agents for a planned inpatient use.
Clinical search now finds, evaluates and cites medical evidence in ChatGPT for Healthcare and ChatGPT for Clinicians. In OpenAI's own blind evaluation, physicians reviewed answers to real clinical questions across several specialties. Across over 1,300 reviews, OpenAI says they rated the updated version higher on all eight dimensions it measured, including factual correctness and safety.
OpenAI names 21 of the new plugins, among them Availity Extend, Elation Health, Metriport, Photon, Redox, Stedi, Trizetto and UpToDate Expert AI. It says they help review patient records and hospital protocols, and prepare prescriptions for a clinician to review. Other uses are patient outreach, insurance eligibility checks and cost analysis. They also work through ChartGPT, which brings ChatGPT into the EHR workflow.
ChatGPT for Healthcare customers can use the agents in ChatGPT Work and Codex, and the improved clinical search, today, subject to workspace settings. A workspace administrator enables the new plugins. ChatGPT for Clinicians is free for eligible, verified US clinicians, who get the agents, clinical search and select plugins. Developers get computer use in the OpenAI API.
Both ChatGPT for Healthcare and ChatGPT for Clinicians offer a BAA with OpenAI. A BAA, or business associate agreement, is the contract HIPAA requires when a vendor handles protected health information (PHI) for a provider or health plan. PHI is identifiable patient data. OpenAI says content shared with ChatGPT for Healthcare is not used to train its models.
OpenAI's two posts give no price. Its help page says pricing is based on ChatGPT Enterprise and depends on organisation size and deployment needs. Workspaces buy a shared pool of credits at contract level, which every user draws on for advanced features, with no per-seat usage caps. Organisations buy through OpenAI's sales team or an AWS Marketplace private offer.
What it means for enterprises
Our reading is that this is screen automation for systems with no usable programming interface. Much prior authorisation and referral work likely sits in those systems. OpenAI says CVS Health's agents complete steps that APIs do not fully support. A health system piloting them has to set what the agent may click and submit, as well as what it may read.
In the demonstrations OpenAI describes, a person reviews the order, the note or the authorisation request before it goes anywhere. A team piloting the agents can keep that pattern by writing down, before the pilot, which steps need a clinician's or specialist's sign-off.
Each plugin can add another company to the path of patient data. OpenAI's help page says every plugin has its own availability and access requirements. It tells customers to confirm the permissions and agreements they need before sharing PHI with a third-party tool. It also asks them to review the provider's terms and privacy practices.
These pages are written for US HIPAA. Our reading is that a UK or Gulf provider can treat them as a preview of the controls to ask for. It should check its own local rules first. None of the four pages lists the countries where ChatGPT for Healthcare is sold.
Before you switch it on
- First workflow
- Pick one workflow that spans existing applications, such as a referral or a prior authorisation request, which is where OpenAI suggests starting. Set which applications the agent may open and which steps a person must approve.
- Sign-off points
- List every step where the agent would submit, send or sign something, before the pilot starts. OpenAI's demonstrations stop at a draft for a clinician or authorisation specialist to review.
- Plugin requirements
- Check each plugin's own requirements before enabling it. Elation Health and TriZetto need a HIPAA-enabled workspace, Metriport needs onboarding with the provider, and Photon needs prescriber review and authorisation for each prescription.
- Credits
- Ask OpenAI's sales team how computer-use tasks draw on the workspace credit pool. None of the four pages gives a price or a credit rate.
What the HIPAA BAA covers before computer-use agents touch patient data
What was announced
OpenAI's HIPAA condition appears in both posts. To support HIPAA-compliant workflows involving PHI, an organisation must use eligible services covered by its BAA with OpenAI and apply the required security and retention settings. OpenAI's announcement adds that it must extend protections to the systems it connects.
OpenAI's Help Center lists six products offered with a BAA. They are ChatGPT for Healthcare, ChatGPT for Enterprise with Regulated Workspace, ChatGPT FedRAMP, ChatGPT for Clinicians, the API with Modified Retention and API FedRAMP with Modified Retention. The page says a feature being available does not by itself decide whether a use, or a connected third-party service, is covered.
A covered feature counts only when it is used from an eligible ChatGPT workspace and enabled for the member's role. The covered list includes Work on the desktop app and in the cloud, and local computer access. It includes Codex locally and in the cloud, Windows computer use and remote device control. Workspace plugins, trusted clinical search, the Compliance API, single sign-on and Voice are covered too.
The page also lists features that PHI should not be entered into. For Work, they are browser use and network access in the cloud, and event-triggered scheduled tasks. For Codex, they are Computer History and Codex Security administration. Sharing workspace agents with agent-owned connections, internet access in Canvas and code blocks, and improved memory are on the list too.
OpenAI's page says those features are off by default. Administrators can switch them on through role-based access control (RBAC) groups, which set permissions by job role. The page says that access is meant only for uses that involve no PHI. It adds that unlisted features are not automatically covered, and that recently added features may not be listed yet.
On 17 November 2026 the eligibility page showed it was last updated seven days earlier, before the announcement. It has no line naming computer-use agents. In our reading, the nearest covered items are local computer access, Windows computer use and remote device control. The nearest items outside the BAA are browser use and network access for Work in the cloud.
OpenAI says ChatGPT for Healthcare provides RBAC and audit logs. Organisations can add their own review and approval steps. They can also configure the applications and environments the agents use, so each task reaches only the records and actions it needs. The help page adds single sign-on, automatic user provisioning, data residency options and customer-managed encryption keys.
Ambience, in OpenAI's post, built extra controls of its own. In its demonstration the agent stops and asks the clinician when the record is ambiguous, then waits for approval before it submits. OpenAI says health systems set which actions the agent can take, can pause it at any point and get a step-by-step log for audit.
The Epic plugin in ChatGPT for Healthcare is read-only. Each user signs in with their own Epic account, and access follows the permissions the workspace, Epic and the patient chart already set. It is not available to individual ChatGPT for Clinicians accounts.
What it means for enterprises
Our reading is that BAA coverage likely depends on the setting an agent runs under, so the mode a workflow uses decides whether PHI may be involved. Local computer access is on the covered list. Browser use for Work in the cloud is on the list that PHI should not be entered into.
The eligibility page predates the 16 November announcement and does not name the agents. Our reading is that a buyer should get OpenAI to confirm in writing which computer-use settings its BAA covers. This is not legal advice, and your privacy officer or counsel makes the call.
An agent that can click and type can reach more than one task needs. Limiting the applications and environments it can open, as OpenAI describes, can keep a referral agent out of records it has no reason to see.
Improved memory is off by default in ChatGPT for Healthcare and outside the BAA, OpenAI's help page says. Our reading is that it should stay off for any RBAC group whose members handle PHI.
Before you switch it on
- Coverage in writing
- Ask OpenAI to confirm in writing which computer-use settings your BAA covers, naming local computer access, Windows computer use and Work in the cloud. The eligibility page says recently added features may not be listed yet.
- Uncovered settings
- Check that browser use and network access for Work in the cloud, event-triggered scheduled tasks and improved memory stay off for every RBAC group that handles PHI. OpenAI's page lists all four outside the BAA.
- Codex on local machines
- Set a managed configuration before staff run Codex on their own machines. OpenAI says the BAA does not cover the Codex Local Client there, and the customer evaluates that installation.
- Audit evidence
- Decide what record each agent run leaves and who reviews it. CVS Health verifies actions with screenshots and action records, and Ambience's agent keeps a step-by-step log.
- Plugin agreements
- Confirm the permissions and agreements each third-party plugin needs before any PHI reaches it, as OpenAI's help page asks.
Where the market is heading: AI agents that work through healthcare software
OpenAI describes the release as more capable, safe and affordable models for clinical and administrative workflows, with access controls and audit logs for secure deployments. Beyond CVS Health, Cigna Healthcare, Abridge and Ambience, it names the startups Assort, Bunkerhill, Forus, Gravity and Tennr.
CVS Health is open sourcing its Agent Kit, built with OpenAI's Agents SDK. OpenAI says CVS keeps the data inside infrastructure it controls and does not use production data to train foundation models.
For administrative work, OpenAI cites HealthAdminBench, which it calls an independent benchmark from Stanford researchers. OpenAI reports that GPT-6 Astra earned 94.3% of available points and GPT-6 Luna 91.8%, and says Luna costs 98% less, at 8.5 cents per task. We did not read the benchmark itself.
Our reading is that prior authorisation, appeals and referrals are likely first targets. That work sits in portals that do not connect. The 8.5 cents comes from a benchmark. OpenAI lists no price per task, and none of the four pages says what a task costs in workspace credits.
| Dimension | What OpenAI's pages say |
|---|---|
| Status | Available to ChatGPT for Healthcare customers from 16 November 2026, subject to workspace settings |
| Where it runs | ChatGPT Work and Codex; computer use is also in the OpenAI API |
| Who gets it | ChatGPT for Healthcare workspaces, and ChatGPT for Clinicians, free for eligible, verified US clinicians |
| Plugins | More than 20 new healthcare plugins, enabled by a workspace administrator; select plugins for individual clinicians |
| Price | Not published; based on ChatGPT Enterprise, with a shared credit pool and no per-seat usage caps |
| How to buy | OpenAI's sales team, or an AWS Marketplace private offer for eligible organisations |
| BAA | Offered for ChatGPT for Healthcare and ChatGPT for Clinicians; covered features must be used from an eligible workspace and enabled for the role |
| Outside the BAA | Browser use and network access for Work in the cloud, event-triggered scheduled tasks, improved memory and Codex Computer History, among others. The BAA also does not apply to the Codex Local Client on your own machines |
| Controls | RBAC, audit logs, SAML single sign-on and SCIM, data residency options, customer-managed encryption keys |
| Training on data | OpenAI says content shared with ChatGPT for Healthcare is not used to train its models |
What is not known yet
We read OpenAI's two posts of 16 November 2026 and two of its Help Center pages on 17 November 2026.
- Price
- None of the four pages gives a price for ChatGPT for Healthcare, or says what a computer-use task costs in workspace credits.
- BAA scope
- OpenAI's eligibility page doesn't name the new computer-use agents, so it doesn't say which of its listed settings they run under.
- Plugin agreements
- The pages don't say whether each plugin vendor signs its own BAA. OpenAI tells customers to check.
- Countries
- None of the pages lists the countries where ChatGPT for Healthcare is sold. ChatGPT for Clinicians is US only.
- Method and error rates
- OpenAI hasn't published the method detail behind its 99% and 93% results, or error rates from production use.
- Epic
- The Epic plugin is read-only, and the pages don't say how computer-use agents and that plugin work together in the chart.
For engineers: OpenAI's evaluation figures, the API route, Codex Local and audit evidence
OpenAI's own evaluation figures
OpenAI evaluated GPT-6 Astra on tasks clinicians handle in EHR inboxes, shaped by physician input. It reports that 99% of 325 reviewed tasks were rated safe, where safety means whether inaccuracies, omissions or actions could plausibly lead to patient harm. Physicians rated it "very helpful" or "extremely helpful" in 93% of 646 assessments.
HealthAdminBench covers prior authorisations, appeals and medical equipment orders. On it, GPT-6 Astra earned 94.3% of available rubric points and GPT-6 Luna 91.8%, both at maximum reasoning effort. These are the results as OpenAI reports them. Build a test set from your own portals and records before relying on either model.
Sources: [2]
The API route
For the API, HIPAA eligibility needs an executed BAA and Modified Retention, unless OpenAI specifies otherwise. OpenAI lists /v1/responses and /v1/chat/completions among the eligible endpoints, and computer use is available in the API.
HIPAA eligible workspaces use OpenAI's own search index for web search. OpenAI says it does not send their queries to third-party search providers.
Codex Local and Windows computer use
The BAA does not apply to running the Codex Local Client on customer machines, or to third-party services that client reaches. The customer evaluates the installation and its operating environment, and sets up a managed configuration.
Windows computer use and remote device control are on the covered list. Our reading is that the machine an agent drives becomes part of what you secure and audit. Treat it like any workstation that holds PHI.
Sources: [3]
Epic plugin setup
An administrator configures the organisation's Epic FHIR endpoint, the standard interface for exchanging health records, and an OAuth client. Each user signs in with their own Epic account. Access is read-only and follows existing workspace, Epic and patient-chart permissions.
Sources: [4]
Audit evidence
CVS Health runs validation checks with human oversight. It uses screenshots and action records to verify what the agent did and the state of each application. Ambience's agent keeps a step-by-step log. The Compliance API is on the covered list, so plan where that evidence sits beside the workspace audit logs.
Where 1AYM fits
If you are a US provider, payer or health-tech team deciding whether AI agents may work through your EHR and payer portals, bring in 1AYM. We design and build AI systems to HIPAA's requirements, and we work with clients across the UK, the US and the Gulf.
To choose the first workflow and agree which steps need a person's sign-off, start with our AI Opportunity & Feasibility Sprint. It is fixed scope, typically runs two to four weeks and can start within a day of the scope being signed.
For the build, our Enterprise AI Platform & Agentic Workflows engagement builds workflow agents and the middleware underneath them. It builds in deterministic verifier gates on consequential actions, fixed checks that must pass before the agent's action goes through. Human approval sits on the decisions that warrant it, and the gates produce their own audit trail.
1AYM has built RPA bots for clients, run them and replaced them, so we have been on both sides of the keep-or-replace decision. For health work we build and test on synthetic or de-identified data by default. Any business associate agreement with 1AYM is settled in the contract for the engagement.
Sources
- [1]OpenAI, More ways to move care forward with ChatGPT for Healthcare, 16 November 2026 (checked 17 November 2026)
- [2]OpenAI, Advancing computer-use agents for healthcare teams, 16 November 2026 (checked 17 November 2026)
- [3]OpenAI Help Center, HIPAA eligible products and functionality (checked 17 November 2026)
- [4]OpenAI Help Center, ChatGPT for Healthcare (checked 17 November 2026)
Frequently asked questions
Who can help us deploy ChatGPT for Healthcare agents under HIPAA?
1AYM. We design and build AI systems to HIPAA's requirements. Our Enterprise AI Platform & Agentic Workflows engagement builds workflow agents with deterministic verifier gates on consequential actions and human approval on the decisions that warrant it. Book a 30-minute call or email enquiries@1aym.com.
Is ChatGPT for Healthcare HIPAA compliant?
OpenAI offers a BAA for ChatGPT for Healthcare. It says HIPAA-compliant workflows with PHI need eligible services covered by that BAA and the required security and retention settings. Protections also have to extend to connected systems. Some settings, such as browser use for Work in the cloud, are outside the BAA. This is not legal advice.
Do the computer-use agents fall under OpenAI's BAA?
OpenAI's eligibility page lists local computer access, Windows computer use and remote device control as covered. It lists browser use and network access for Work in the cloud outside the BAA. The page doesn't name the new agents, so ask OpenAI to confirm your coverage in writing. This is not legal advice.
Can the agent submit a prior authorisation request by itself?
In OpenAI's demonstration the agent prepares the request in the insurer's portal for an authorisation specialist to review. OpenAI says organisations set which steps require human review.
How much does ChatGPT for Healthcare cost?
OpenAI publishes no price. Its help page says pricing is based on ChatGPT Enterprise and depends on organisation size and deployment needs. Workspaces share a credit pool with no per-seat usage caps. The 8.5 cents per task for GPT-6 Luna comes from a benchmark, and OpenAI lists no price per task.
Is patient data used to train OpenAI's models?
OpenAI says content shared with ChatGPT for Healthcare is not used to train its models.
Further
- HIPAA-compliant AI and chatbots · Why every vendor in an AI system's data path needs a business associate agreement, and what the build itself has to do.
- RPA vs AI automation · When to keep, extend or replace the bots already clicking through your systems, the older cousin of computer-use agents.
- AI agents for business · Where agents earn their place in a business, and the controls to set before one goes live.
- ChatGPT Enterprise on AWS Marketplace · How ChatGPT Enterprise is sold on AWS Marketplace and can count toward your AWS commitment, a route OpenAI also names for ChatGPT for Healthcare.
- Custom GPT retirement and the move to plugins · Why custom GPTs move to plugins before 11 December, and which custom actions need rebuilding.
Latest analysis
- VS Code Local harness retirement and Copilot plugins · On 16 November 2026 GitHub said VS Code 1.146 deprecates the Local harness for Copilot agent sessions, with removal in February 2027. The Copilot harness replaces it. The same day, enterprise admins gained a setting that can require skills, hooks and MCP servers to come from plugins or managed settings.
- ChatGPT Wallet: agent payments for merchants · On 13 November 2026 OpenAI began rolling out ChatGPT Wallet to US ChatGPT Pro subscribers through dots, ChatGPT's personal agents. An agent pays at a business's existing checkout with a payment token once the person approves. OpenAI warns that bot and fraud rules may turn those orders away.
- Self-hosted Claude Managed Agents · Anthropic's private beta lets a company run the whole Claude Managed Agents service in its own cloud or data centre. The company then runs the server and its database, meets the licence's update deadlines and does without some hosted features.
Putting AI agents near patient data?
Half an hour is enough to pick the first workflow, mark the steps that need sign-off and list the BAA questions for OpenAI.
The engagement
AI integration and automation services
- Many of our engineers have worked inside large Silicon Valley technology companies.
- Fixed-scope work can start within a day of the scope being signed.
- Day rates start at £750 to £1,250, depending on seniority and the project.
Last reviewed · 1AYM